未验证 提交 69510524 编写于 作者: L LielinJiang 提交者: GitHub

add default conv init (#29092)

上级 fea0e294
...@@ -123,8 +123,17 @@ class _ConvNd(layers.Layer): ...@@ -123,8 +123,17 @@ class _ConvNd(layers.Layer):
filter_shape = [out_channels, in_channels // groups filter_shape = [out_channels, in_channels // groups
] + self._kernel_size ] + self._kernel_size
def _get_default_param_initializer():
if transposed:
return None
filter_elem_num = np.prod(self._kernel_size) * self._in_channels
std = (2.0 / filter_elem_num)**0.5
return Normal(0.0, std, 0)
self.weight = self.create_parameter( self.weight = self.create_parameter(
shape=filter_shape, attr=self._param_attr) shape=filter_shape,
attr=self._param_attr,
default_initializer=_get_default_param_initializer())
self.bias = self.create_parameter( self.bias = self.create_parameter(
attr=self._bias_attr, shape=[self._out_channels], is_bias=True) attr=self._bias_attr, shape=[self._out_channels], is_bias=True)
......
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册